Write Drunk, edit sober (with ChatGPT)

If your media consumption is anything like mine, your feed this week has been flooded with hot takes on a new bit of research out of MIT’s Media Lab – research which, on the surface, suggests that letting ChatGPT write for you impedes your cognitive function.

‍

The study being referenced was conducted in Boston with 54 subjects aged 18-38, divided up into three groups. All of them were asked to write essays, one group using ChatGPT, one using only their brain, and one using a search engine. It measured brain engagement and function throughout the process. On the surface, the findings suggest that participants using AI from beginning to end ‘consistently underperformed’ on neural, linguistic, and behavioural levels. Their brains weren’t firing as hard, essays weren’t as good, and once they’d written them, they could barely quote them back.

‍

But what I found fascinating about the study wasn’t the endless churn of headlines about ChatGPT ‘making us dumber’ or ‘giving us brain rot’ (yawn)… it’s what happened when the ‘brain only’ group got given access to AI halfway through the study, after they’d already used their actual brains. Because that cohort’s brains were firing on all cylinders.

‍

The findings didn’t feel anti-AI – they felt anti-AI as our first port of call. A reminder that when we use these tools is just as important as how we use them. So, read this as my fairly longwinded defence of the wiggly tangential first bit of every brief - speaking to people, scrolling niche Reddit threads, going to bad exhibitions, reading deeply AND sparring with our bot friends.

‍

As naff and misattributed as ‘write drunk, edit sober’ might be, there’s something in the idea that first drafts are good when they’re flawed and human and full of mismatched threads of thought. That’s the whole point.

‍

We get the most out of AI when we do a bit of thinking first. Why let the robots do the fun bit?

‍Read the research for yourself.